A Treebank-based Investigation of IPP-triggering Verbs in Dutch

phrases are generated by a shallow parsing model (hidden Markov model) trained
on the Le Monde corpus [1]. Finally, the monolingual treebanks are exported in
the TIGER-XML format [6]. More details about the annotation of the German
treebank can be found in [14], whereas the French annotation is described in [5].
We use the Bleualign algorithm [13] to align the sentences across both monolingual
treebanks. Our alignment convention was to discard the automatic many-tomany
word alignments for the purpose of increasing the precision. Subsequently,
a human annotator checked and, when required, corrected the remaining word and
sentence alignments and then added the phrase alignments. Finally, the alignment
file is available in XML format, as the following snippet shows:
This says that node 18 in sentence 225 of the German treebank (de) is aligned
with node 16 in sentence 231 of the French treebank (fr). The node identifiers refer
to the IDs in the TIGER-XML treebanks. The alignment is labeled as good when
the linked phrases represent exact translations and as fuzzy in case of approximate
correspondences.
3 The Evaluation Tool: DELiC4MT
DELiC4MT is an open-source tool that performs diagnostic evaluation of MT systems
over user-defined linguistically-motivated constructions, also called checkpoints.
This term was introduced by Zhou et al. [15] and refers to either lexical
elements or grammatical constructions, such as ambiguous words, noun phrases,
verb-object collocations etc. The experiments reported in this paper follow the
workflow proposed by Naskar et al. [9], due to its option to integrate new language
pairs.
First the texts are PoS-tagged and exported in the KYOTO Annotation Format
(KAF)[2]. This scheme facilitates the inspection of the terms in the sentences and
thus querying for specific features, such as PoS sequences. Figure 1 depicts the
KAF annotation of the German phrase den ersten Gipfel (EN: the first peak) and
its French equivalent le premier sommet.
The linguistic checkpoints are subsequently defined in the so-called kybot profiles.
A kybot profile starts with the declaration of the involved variables and the
relations among them and ends specifying which attributes of the matched terms
should be exported. For example, figure 2 depicts the kybot profile for a nominal
group consisting of a determiner, an adjective and a noun. Moreover, the constituent
terms have to be consecutive. Once defined, the kybot profile is run over
the source language KAF files and the matched terms (with the specified attributes)
are stored in a separate XML file.
147